There’s a date circled on every compliance calendar in Europe right now: 2 August 2026. That’s when the EU AI Act’s transparency rules of Article 50 start to bite.
If your business builds, deploys, or even quietly relies on AI that talks to people, generates content, or reads faces, this is the part of the AI Act you can’t wave away as someone else’s problem. Here’s the short version, and why the details are worth a closer look.
Different use-cases, different rules
Article 50 isn’t one rule. It’s four, each aimed at a different use-case:
- Chatbots and AI agents must tell people they’re not human. If your customer-service bot, voice assistant, or AI companion is designed to feel human, “surprise, I’m a robot” needs to be built in, not buried in the terms and conditions. The way this message is conveyed, differs based on the circumstances and the (intended) audience.
- AI-generated content (images, audio, video, text) has to be machine-readable and marked as synthetic, so it can be detected downstream. Think watermarks, metadata, provenance signals.
- Emotion recognition and biometric categorisation systems require you to tell people they’re being analysed.
- Deepfakes and AI-written news must be clearly labelled as artificial when they resemble real people or inform the public on matters of public interest.
Simple enough on paper. The complications live in the fine print and each of the different actors (the Providers of the AI System vs. the Deployers of the AI System) each have a different role to play.
The traps hiding in the details
A few things may catch people off guard:
Chatbots and AI agents must tell people they’re not human. You can skip the chatbot disclosure only if it’s genuinely obvious to a reasonable person that they’re talking to AI. But ‘obvious’ is a high standard. The Commission reads this narrowly, and when children, elderly users, or vulnerable groups might be in the audience, the bar is set even higher. A friendly human-sounding voice assistant almost certainly doesn’t get the free pass. Do You want to go back to a robot-voice?
Your Christmas card is fine. Your political meme is not. Personal, non-professional use is carved out from the required AI labeling, so the deepfake you make of your own family for a holiday card doesn’t need a label. But post a deepfake of your mayor to criticise a local policy, and you’re suddenly outside the exemption. The line between “personal” and “public interest” is thinner than most people assume.
“AI helped me edit” might still count. Minor cleanup, such as grammar, cropping, colour correction is exempt from the labeling requirements. But translations, summaries, adding or removing objects, or anything that changes meaning? That’s substantive manipulation, and it needs marking. The gap between “assistive editing” and “content generation” is where a lot of real-world tools live.
Open source doesn’t get you out. Released your model under a free and open-source licence? The transparency obligations still apply. Full stop.
One deadline just moved (a little)
There’s a small mercy. Under the AI Act “Omnibus” changes agreed in May 2026, generative AI systems already on the market before 2 August 2026 get a short grace period until 2 December 2026 to meet the marking-and-detection requirements. The other three obligations? No transition. They apply from day one.